Non-Destructive and Real-Time Discrimination of Normal and Frozen-Thawed Beef Based on a Novel Deep Learning Model.

Area of Science:

• Food Science
• Computer Vision
• Artificial Intelligence

Background:

• Ensuring food safety requires reliable methods to differentiate fresh and frozen-thawed beef.
• Current methods may be destructive or lack real-time capabilities, posing challenges for quality assessment.

Purpose of the Study:

• To develop a novel, non-destructive, and real-time You Only Look Once for normal and frozen-thawed beef discrimination (YOLO-NF) model.
• To enhance the YOLO-NF model's performance using attention mechanisms for improved beef discrimination.

Main Methods:

• Utilized deep learning techniques, specifically the You Only Look Once (YOLO) architecture, enhanced with Simple, Parameter-free Attention Module (SimAM) and Squeeze-and-Excitation (SE) mechanisms.
• Trained and validated the model on 1200 beef samples (images captured by CCD camera), with data augmentation applied to the training set.
• Employed Gradient-weighted Class Activation Mapping (Grad-CAM) for model interpretability and deployed the model on a web interface for user accessibility.

Main Results:

• The YOLO-NF model achieved high performance metrics on the test set: 95.5% precision, 95.2% recall, 95.3% F1-Score, and 98.6% mean average precision (mAP).
• Demonstrated superior performance compared to existing YOLOv7, YOLOv5, and YOLOv8 models in beef discrimination tasks.
• Achieved a rapid discrimination time of 0.94 seconds per image on a local server, confirming its real-time processing capability.

Conclusions:

• The developed YOLO-NF model offers a highly accurate and efficient solution for non-destructive, real-time discrimination between fresh and frozen-thawed beef.
• This technique presents a promising advancement for meat quality assessment within food safety monitoring systems.
• The model's interpretability and web deployment enhance its practical applicability in the food industry.